It felt surreal to write this at the time. After four years at Wuhan University and an internship at Microsoft Research Asia, I had begun my Ph.D. at the University of Hawai‘i at Mānoa.
Why privacy and LLM agents?
During my time at Microsoft Research Asia, I worked on understanding how LLM-powered agents handle privacy in realistic settings. The more I studied the problem, the clearer it became that there is a meaningful gap between what models can say about privacy norms and how they behave when deployed as agents.
That judgment–action gap became the seed of my subsequent work on contextual privacy and agent evaluation.
What I was working on
My research then focused on multimodal contextual integrity for LLM and VLM agents. Contextual integrity, introduced by Helen Nissenbaum, is a framework for thinking about privacy as appropriate information flow. I was interested in extending it to settings where agents reason over images, text, and structured data together.
Privacy is not only about secrecy. It is about the appropriate flow of information in context.
This archive is where I may share research notes and reflections over time. Thanks for reading.